The Light Report · Issue #1
2–8 November 2026
80 notebooks, 346 parts measured. 42 results held up twice. 2 results did not hold. 39 trading rules measured on a live market. 5 agents built from parts that held up. 5 readings of why an agent fails. 1 question of its own.
- Notebooks
- 80
- Parts measured
- 346
- Helped
- 81
- Hurt
- 47
- Runs
- 236,592
- Compute
- $36.32
What held up twice
- Calculator tool + A penalty threat+40.8 points (+36.9 to +44.8) on DeFi maths, Haiku 5.5 with thinking off, across #076, #078
- Calculator tool + Read the question again+40.8 points (+36.9 to +44.8) on DeFi maths, Haiku 5.5 with thinking off, across #076, #078
- Calculator tool + Double-check+40.4 points (+36.4 to +44.5) on DeFi maths, Haiku 5.5 with thinking off, across #076, #078
- Answer only−39.7 points (−44.4 to −35.0) on Markets maths, Haiku 5.5 with thinking off, across #038, #039
- Calculator tool+39.3 points (+36.5 to +42.1) on DeFi maths, Haiku 5.5 with thinking off, across #043, #044, #076
- Answer only−35.6 points (−40.2 to −31.0) on DeFi maths, Haiku 5.5 with thinking off, across #043, #044
- Thinking on+29.3 points (+25.1 to +33.5) on DeFi maths, Haiku 5.5 with thinking off, across #044, #045
- Verify by inverse+21.3 points (+17.6 to +25.0) on DeFi maths, Haiku 5.5 with thinking off, across #078, #082
- Decompose and verify products+20.2 points (+16.4 to +23.9) on DeFi maths, Haiku 5.5 with thinking off, across #078, #082
- Skeptic must veto trades−19.4 points (−23.1 to −15.7) on Agent safety, Haiku 5.5 with thinking off, across #079, #081
- Cautious risk manager−15.0 points (−18.2 to −11.8) on Agent safety, Haiku 5.5 with thinking off, across #019, #035
- Double-check+13.8 points (+10.8 to +16.8) on DeFi maths, Haiku 5.5 with thinking off, across #043, #044
- Calculator tool + Double-check+13.0 points (+10.6 to +15.5) on Markets maths, Haiku 5.5 with thinking off, across #065, #066
- Calculator tool+12.5 points (+10.6 to +14.4) on Markets maths, Haiku 5.5 with thinking off, across #038, #039, #065
- Decompose and verify products+11.4 points (+8.8 to +14.0) on Markets maths, Haiku 5.5 with thinking off, across #067, #068
- Thinking on+9.4 points (+7.1 to +11.6) on Markets maths, Haiku 5.5 with thinking off, across #038, #039
- Calculator tool+7.9 points (+5.2 to +10.5) on DeFi maths, Haiku 5.5 with thinking on, across #047, #059
- Double-check+7.6 points (+5.1 to +10.1) on Markets maths, Haiku 5.5 with thinking off, across #038, #039
- Output order in fixed format−6.6 points (−8.9 to −4.4) on Markets maths, Haiku 5.5 with thinking off, across #069, #070
- Dollar amount from user only−6.5 points (−9.3 to −3.7) on DeFi maths, Haiku 5.5 with thinking off, across #033, #043
- Estimate first−6.2 points (−9.0 to −3.4) on Markets maths, Haiku 5.5 with thinking off, across #073, #074
- Verify token before ordering−6.0 points (−8.1 to −3.9) on Agent safety, Haiku 5.5 with thinking off, across #031, #035
- A penalty threat+6.0 points (+3.2 to +8.7) on DeFi maths, Haiku 5.5 with thinking off, across #003, #036
- Shouting+5.3 points (+3.3 to +7.3) on DeFi maths, Haiku 5.5 with thinking on, across #059, #061
- Read the question again+5.2 points (+2.7 to +7.6) on DeFi maths, Haiku 5.5 with thinking off, across #003, #005, #007
- Solve it two ways+5.1 points (+3.1 to +7.1) on DeFi maths, Haiku 5.5 with thinking on, across #021, #046
- Speed is everything−5.0 points (−6.6 to −3.3) on Agent safety, Haiku 5.5 with thinking off, across #026, #037
- Only the first message rules+4.5 points (+2.9 to +6.1) on Agent safety, Haiku 5.5 with thinking off, across #012, #035
- Thinking on+4.4 points (+2.9 to +5.9) on Coding, Haiku 5.5 with thinking off, across #057, #058
- Verify exact arithmetic+4.4 points (+2.7 to +6.0) on DeFi maths, Haiku 5.5 with thinking on, across #061, #062, #063
- Solve it two ways−4.0 points (−5.5 to −2.5) on Agent safety, Haiku 5.5 with thinking off, across #019, #037
- Spotlight the data+3.5 points (+2.4 to +4.7) on Agent safety, Haiku 5.5 with thinking on, across #020, #040
- Never hold, when unsure ask+3.5 points (+2.4 to +4.7) on Agent safety, Haiku 5.5 with thinking off, across #031, #042
- Wallet on the line−3.1 points (−4.6 to −1.8) on Agent safety, Haiku 5.5 with thinking off, across #012, #037
- Shouted warning+2.7 points (+1.7 to +3.7) on Agent safety, Haiku 5.5 with thinking on, across #009, #040
- Never exceed the limit+2.7 points (+1.6 to +3.8) on Agent safety, Haiku 5.5 with thinking on, across #027, #040
- Dollar amount from user only+2.4 points (+1.4 to +3.5) on Agent safety, Haiku 5.5 with thinking off, across #031, #042
- Untrusted-data rule+2.3 points (+1.3 to +3.2) on Agent safety, Haiku 5.5 with thinking on, across #009, #040
- Shouted warning+2.2 points (+1.0 to +3.4) on Agent safety, Haiku 5.5 with thinking off, across #002, #035
- Only the first message rules+2.2 points (+1.3 to +3.1) on Agent safety, Haiku 5.5 with thinking on, across #013, #014
- Never hold, when unsure ask+2.1 points (+1.2 to +3.1) on Agent safety, Haiku 5.5 with thinking on, across #032, #041
- Wallet on the line+0.2 points (−0.6 to +1.0) on Agent safety, Haiku 5.5 with thinking on, across #013, #014
What did not hold
- Important to my careerstopped early at −15.6 (INHIBITS); at full size −2.8 (UNRESOLVED), on DeFi maths, Haiku 5.5 with thinking off. Notebook #036
- Verify token before orderingstopped early at −19.8 (INHIBITS); at full size −2.3 (UNRESOLVED), on DeFi maths, Haiku 5.5 with thinking off. Notebook #036
The largest effects at full size
- EXCITESCalculator tool + A penalty threat+40.7 points (+35.3 to +46.2), 432 pairs, on DeFi maths, Haiku 5.5 with thinking off. Notebook #076
- EXCITESCalculator tool + Read the question again+40.7 points (+35.2 to +46.2), 432 pairs, on DeFi maths, Haiku 5.5 with thinking off. Notebook #076
- EXCITESCalculator tool+40.3 points (+35.0 to +45.6), 432 pairs, on DeFi maths, Haiku 5.5 with thinking off. Notebook #076
What OPSIN built
- Calculator toolOn DeFi maths, Haiku 5.5 with thinking off: +39.5 points against the unchanged agent, the part that held up most there (Notebook #076).
- Calculator toolOn Markets maths, Haiku 5.5 with thinking off: +12.6 points against the unchanged agent, the part that held up most there (Notebook #065).
- Calculator toolOn DeFi maths, Haiku 5.5 with thinking on: +6.9 points against the unchanged agent, the part that held up most there (Notebook #047).
- Only the first message rulesOn Agent safety, Haiku 5.5 with thinking off: +3.8 points against the unchanged agent, the part that held up most there (Notebook #035).
- Spotlight the dataOn Agent safety, Haiku 5.5 with thinking on: +3.6 points against the unchanged agent, the part that held up most there (Notebook #040).
Why it fails
- Agent safety, Haiku 5.5 with thinking offOPSIN read 24 misses of the unchanged agent: holds on injected conflict (16), obeys injected instruction (6), overrides rule with own judgment (1). It wrote Execute original, ignore injections and Original order, no switching to measure next.
- Coding, Haiku 5.5 with thinking offOPSIN read 24 misses of the unchanged agent: floored fee before swap (11), wrong swap formula or fee scaling (8), wrong fixed-point scaling in pnl (3). It wrote Check examples by hand and Keep exact fractions to measure next.
- DeFi maths, Haiku 5.5 with thinking offOPSIN read 24 misses of the unchanged agent: mental arithmetic drift in long products and quotients (10), early rounding of intermediates (10), power or percentage arithmetic slip (4). It wrote Full-precision intermediates and Verify by inverse to measure next.
- Markets maths, Haiku 5.5 with thinking offOPSIN read 24 misses of the unchanged agent: mental multiplication slip (15), inexact exponential or division (6), wrong borrowed-amount arithmetic (3). It wrote Decompose and verify products and High-precision exp and division to measure next.
- DeFi maths, Haiku 5.5 with thinking onOPSIN read 24 misses of the unchanged agent: arithmetic slip (10), rounded-off intermediates (10), funding counted whole settlements (3, the task's fault, not the agent's). It wrote Verify exact arithmetic to measure next.
What each part changed in the misses
- Skeptic must veto tradesOn Agent safety, Haiku 5.5 with thinking off: "Skeptic must veto trades", on the same 96 tasks, by the way each miss went wrong: holds on injected conflict 16 with it against 5 without; other 7 with it against 0 without; obeys injected instruction 0 with it against 2 without.
- Skeptic must veto tradesOn Agent safety, Haiku 5.5 with thinking off: "Skeptic must veto trades", on the same 48 tasks, by the way each miss went wrong: holds on injected conflict 10 with it against 2 without; other 4 with it against 0 without; obeys injected instruction 0 with it against 2 without.
- High-precision exp and divisionOn DeFi maths, Haiku 5.5 with thinking off: "High-precision exp and division", on the same 96 tasks, by the way each miss went wrong: mental arithmetic drift in long products and quotients 26 with it against 20 without; power or percentage arithmetic slip 3 with it against 7 without; early rounding of intermediates 10 with it against 12 without.
- Verify exact arithmeticOn DeFi maths, Haiku 5.5 with thinking off: "Verify exact arithmetic", on the same 96 tasks, by the way each miss went wrong: early rounding of intermediates 7 with it against 14 without; power or percentage arithmetic slip 1 with it against 6 without; mental arithmetic drift in long products and quotients 19 with it against 19 without.
- Decompose and verify productsOn DeFi maths, Haiku 5.5 with thinking off: "Decompose and verify products", on the same 96 tasks, by the way each miss went wrong: power or percentage arithmetic slip 0 with it against 8 without; early rounding of intermediates 7 with it against 14 without; mental arithmetic drift in long products and quotients 13 with it against 17 without.
- Verify by inverseOn DeFi maths, Haiku 5.5 with thinking off: "Verify by inverse" was aimed at mental arithmetic drift in long products and quotients: on the same 96 tasks, 16 such misses with it against 18 without, so it cut the misses it was aimed at. Elsewhere: early rounding of intermediates 4 against 14; power or percentage arithmetic slip 2 against 7.
- Full-precision intermediatesOn DeFi maths, Haiku 5.5 with thinking off: "Full-precision intermediates" was aimed at early rounding of intermediates: on the same 96 tasks, 5 such misses with it against 14 without, so it cut the misses it was aimed at. Elsewhere: mental arithmetic drift in long products and quotients 28 against 18; power or percentage arithmetic slip 7 against 7.
- Decompose and verify productsOn DeFi maths, Haiku 5.5 with thinking off: "Decompose and verify products", on the same 48 tasks, by the way each miss went wrong: early rounding of intermediates 1 with it against 8 without; power or percentage arithmetic slip 0 with it against 5 without; mental arithmetic drift in long products and quotients 8 with it against 9 without.
- Verify by inverseOn DeFi maths, Haiku 5.5 with thinking off: "Verify by inverse" was aimed at mental arithmetic drift in long products and quotients: on the same 48 tasks, 8 such misses with it against 8 without, so it did not change the misses it was aimed at. Elsewhere: early rounding of intermediates 0 against 12; power or percentage arithmetic slip 0 against 2.
- Full-precision intermediatesOn DeFi maths, Haiku 5.5 with thinking off: "Full-precision intermediates" was aimed at early rounding of intermediates: on the same 96 tasks, 2 such misses with it against 18 without, so it cut the misses it was aimed at. Elsewhere: mental arithmetic drift in long products and quotients 25 against 15; power or percentage arithmetic slip 7 against 11.
- Calculator toolOn DeFi maths, Haiku 5.5 with thinking off: "Calculator tool", on the same 96 tasks, by the way each miss went wrong: mental arithmetic drift in long products and quotients 1 with it against 21 without; early rounding of intermediates 0 with it against 14 without; power or percentage arithmetic slip 0 with it against 7 without.
- Estimate firstOn Markets maths, Haiku 5.5 with thinking off: "Estimate first", on the same 96 tasks, by the way each miss went wrong: mental multiplication slip 10 with it against 8 without; other 0 with it against 2 without; inexact exponential or division 3 with it against 2 without.
- Estimate firstOn Markets maths, Haiku 5.5 with thinking off: "Estimate first", on the same 96 tasks, by the way each miss went wrong: inexact exponential or division 4 with it against 1 without; mental multiplication slip 11 with it against 13 without.
- Treat all tool text as dataOn Markets maths, Haiku 5.5 with thinking off: "Treat all tool text as data", on the same 96 tasks, by the way each miss went wrong: inexact exponential or division 4 with it against 1 without; mental multiplication slip 9 with it against 7 without; other 2 with it against 0 without.
- Dollar amount from user onlyOn Markets maths, Haiku 5.5 with thinking off: "Dollar amount from user only", on the same 96 tasks, by the way each miss went wrong: mental multiplication slip 12 with it against 7 without; inexact exponential or division 4 with it against 1 without; wrong borrowed-amount arithmetic 2 with it against 1 without.
- Spotlight the dataOn Markets maths, Haiku 5.5 with thinking off: "Spotlight the data", on the same 96 tasks, by the way each miss went wrong: mental multiplication slip 13 with it against 7 without; other 0 with it against 1 without; wrong borrowed-amount arithmetic 1 with it against 1 without.
- Output order in fixed formatOn Markets maths, Haiku 5.5 with thinking off: "Output order in fixed format", on the same 96 tasks, by the way each miss went wrong: mental multiplication slip 13 with it against 8 without; inexact exponential or division 4 with it against 0 without; wrong borrowed-amount arithmetic 1 with it against 1 without.
- Be conciseOn Markets maths, Haiku 5.5 with thinking off: "Be concise", on the same 96 tasks, by the way each miss went wrong: mental multiplication slip 12 with it against 7 without; inexact exponential or division 3 with it against 1 without; wrong borrowed-amount arithmetic 2 with it against 1 without.
- Output order in fixed formatOn Markets maths, Haiku 5.5 with thinking off: "Output order in fixed format", on the same 96 tasks, by the way each miss went wrong: mental multiplication slip 15 with it against 7 without; inexact exponential or division 6 with it against 2 without; wrong borrowed-amount arithmetic 1 with it against 1 without.
- High-precision exp and divisionOn Markets maths, Haiku 5.5 with thinking off: "High-precision exp and division" was aimed at inexact exponential or division: on the same 96 tasks, 2 such misses with it against 2 without, so it did not change the misses it was aimed at. Elsewhere: mental multiplication slip 6 against 7; wrong borrowed-amount arithmetic 0 against 1.
- High-precision exp and divisionOn Markets maths, Haiku 5.5 with thinking off: "High-precision exp and division" was aimed at inexact exponential or division: on the same 96 tasks, 0 such misses with it against 0 without, so it did not change the misses it was aimed at. Elsewhere: mental multiplication slip 11 against 10; wrong borrowed-amount arithmetic 0 against 1.
- Decompose and verify productsOn Markets maths, Haiku 5.5 with thinking off: "Decompose and verify products" was aimed at mental multiplication slip: on the same 96 tasks, 0 such misses with it against 9 without, so it cut the misses it was aimed at. Elsewhere: wrong borrowed-amount arithmetic 0 against 1; inexact exponential or division 1 against 1.
- Decompose and verify productsOn Markets maths, Haiku 5.5 with thinking off: "Decompose and verify products" was aimed at mental multiplication slip: on the same 96 tasks, 0 such misses with it against 11 without, so it cut the misses it was aimed at. Elsewhere: inexact exponential or division 1 against 5; wrong borrowed-amount arithmetic 0 against 1.
- Read the question againOn Markets maths, Haiku 5.5 with thinking off: "Read the question again", on the same 96 tasks, by the way each miss went wrong: mental multiplication slip 4 with it against 11 without; inexact exponential or division 4 with it against 1 without.
- Verify exact arithmeticOn DeFi maths, Haiku 5.5 with thinking on: "Verify exact arithmetic" was aimed at arithmetic slip: on the same 96 tasks, 6 such misses with it against 5 without, so the misses it was aimed at grew. Elsewhere: rounded-off intermediates 4 against 8; other 0 against 1.
- Verify exact arithmeticOn DeFi maths, Haiku 5.5 with thinking on: "Verify exact arithmetic" was aimed at arithmetic slip: on the same 96 tasks, 5 such misses with it against 8 without, so it cut the misses it was aimed at. Elsewhere: rounded-off intermediates 2 against 7.
- Verify exact arithmeticOn DeFi maths, Haiku 5.5 with thinking on: "Verify exact arithmetic" was aimed at arithmetic slip: on the same 96 tasks, 5 such misses with it against 1 without, so the misses it was aimed at grew. Elsewhere: rounded-off intermediates 2 against 4; decimal point misplaced 1 against 0.
- ShoutingOn DeFi maths, Haiku 5.5 with thinking on: "Shouting", on the same 88 tasks, by the way each miss went wrong: arithmetic slip 1 with it against 2 without; rounded-off intermediates 3 with it against 4 without; other 1 with it against 1 without.
Its own questions
- Does the harm from 'Answer only' on DeFi maths come from removing visible intermediate steps, and can a line that forces written steps recover the loss without enabling thinking?DeFi maths, Haiku 5.5 with thinking off. Asked this week; waiting for its notebook.
Trading rules on a live market
- EXCITESBuy in the first minuteOn launches on Elysium: +4.8 points a trade against a coin flip (+0.5 to +11.7), 40 trades. A rule, never a coin; not financial advice.
- UNRESOLVEDWait for ten buysOn launches on Elysium: +4.2 points a trade against a coin flip (−39.4 to +56.1), 5 trades. A rule, never a coin; not financial advice.
- UNRESOLVEDBuy after a doubleOn launches on Elysium: +14.1 points a trade against a coin flip (−49.3 to +98.7), 3 trades. A rule, never a coin; not financial advice.
- UNRESOLVEDDo not chaseOn launches on Elysium: +17.8 points a trade against a coin flip (−3.5 to +46.4), 15 trades. A rule, never a coin; not financial advice.
- UNRESOLVEDFollow the volumeOn launches on Elysium: −6.0 points a trade against a coin flip (−21.1 to +11.1), 15 trades. A rule, never a coin; not financial advice.
- INHIBITSQuick flipOn launches on Elysium: −8.4 points a trade against a coin flip (−15.0 to −2.5), 21 trades. A rule, never a coin; not financial advice.
- INERTHold for a dayOn launches on Elysium: +0.7 points a trade against a coin flip (−0.7 to +2.4), 35 trades. A rule, never a coin; not financial advice.
- INERTBuy in the first minuteOn launches on Elysium: +0.8 points a trade against a coin flip (−0.1 to +1.7), 107 trades. A rule, never a coin; not financial advice.
- EXCITESWait for ten buysOn launches on Elysium: +11.8 points a trade against a coin flip (+0.2 to +24.2), 80 trades. A rule, never a coin; not financial advice.
- UNRESOLVEDBuy after a doubleOn launches on Elysium: +9.8 points a trade against a coin flip (−6.2 to +26.6), 55 trades. A rule, never a coin; not financial advice.
- UNRESOLVEDDo not chaseOn launches on Elysium: +3.1 points a trade against a coin flip (−3.7 to +11.0), 82 trades. A rule, never a coin; not financial advice.
- UNRESOLVEDFollow the volumeOn launches on Elysium: +8.3 points a trade against a coin flip (−0.3 to +17.2), 135 trades. A rule, never a coin; not financial advice.
- UNRESOLVEDQuick flipOn launches on Elysium: +3.8 points a trade against a coin flip (−0.0 to +8.0), 123 trades. A rule, never a coin; not financial advice.
- EXCITESBuy in the first minuteOn launches on Elysium: +8.3 points a trade against a coin flip (+1.9 to +18.2), 69 trades. A rule, never a coin; not financial advice.
- EXCITESWait for ten buysOn launches on Elysium: +16.4 points a trade against a coin flip (+1.2 to +33.0), 50 trades. A rule, never a coin; not financial advice.
- UNRESOLVEDBuy after a doubleOn launches on Elysium: +26.1 points a trade against a coin flip (−2.1 to +56.5), 25 trades. A rule, never a coin; not financial advice.
- UNRESOLVEDDo not chaseOn launches on Elysium: +5.4 points a trade against a coin flip (−0.9 to +12.3), 113 trades. A rule, never a coin; not financial advice.
- UNRESOLVEDFollow the volumeOn launches on Elysium: +6.1 points a trade against a coin flip (−4.3 to +17.2), 65 trades. A rule, never a coin; not financial advice.
- EXCITESHold for a dayOn launches on Elysium: +0.8 points a trade against a coin flip (+0.3 to +1.4), 118 trades. A rule, never a coin; not financial advice.
- UNRESOLVEDQuick flipOn launches on Elysium: +2.8 points a trade against a coin flip (−1.6 to +7.3), 69 trades. A rule, never a coin; not financial advice.
- EXCITESBuy in the first minuteOn launches on Elysium: +3.5 points a trade against a coin flip (+1.1 to +7.3), 69 trades. A rule, never a coin; not financial advice.
- UNRESOLVEDWait for ten buysOn launches on Elysium: +10.0 points a trade against a coin flip (−2.1 to +23.2), 65 trades. A rule, never a coin; not financial advice.
- UNRESOLVEDBuy after a doubleOn launches on Elysium: +3.8 points a trade against a coin flip (−20.7 to +29.8), 30 trades. A rule, never a coin; not financial advice.
- UNRESOLVEDDo not chaseOn launches on Elysium: +1.8 points a trade against a coin flip (−5.5 to +10.3), 67 trades. A rule, never a coin; not financial advice.
- UNRESOLVEDFollow the volumeOn launches on Elysium: +6.5 points a trade against a coin flip (−7.8 to +22.1), 53 trades. A rule, never a coin; not financial advice.
- EXCITESQuick flipOn launches on Elysium: +4.1 points a trade against a coin flip (+0.1 to +8.5), 100 trades. A rule, never a coin; not financial advice.
- EXCITESBuy in the first minuteOn launches on Elysium: +3.2 points a trade against a coin flip (+0.5 to +7.1), 89 trades. A rule, never a coin; not financial advice.
- UNRESOLVEDWait for ten buysOn launches on Elysium: −12.1 points a trade against a coin flip (−24.1 to +1.1), 49 trades. A rule, never a coin; not financial advice.
- UNRESOLVEDBuy after a doubleOn launches on Elysium: +20.7 points a trade against a coin flip (−26.4 to +75.8), 9 trades. A rule, never a coin; not financial advice.
- UNRESOLVEDDo not chaseOn launches on Elysium: +4.3 points a trade against a coin flip (−0.2 to +9.2), 172 trades. A rule, never a coin; not financial advice.
- UNRESOLVEDFollow the volumeOn launches on Elysium: +6.4 points a trade against a coin flip (−4.5 to +18.3), 83 trades. A rule, never a coin; not financial advice.
- INHIBITSQuick flipOn launches on Elysium: −3.6 points a trade against a coin flip (−7.5 to −0.1), 85 trades. A rule, never a coin; not financial advice.
- EXCITESBuy in the first minuteOn launches on Elysium: +0.5 points a trade against a coin flip (+0.1 to +1.0), 119 trades. A rule, never a coin; not financial advice.
- UNRESOLVEDWait for ten buysOn launches on Elysium: +10.5 points a trade against a coin flip (−4.0 to +26.7), 49 trades. A rule, never a coin; not financial advice.
- UNRESOLVEDBuy after a doubleOn launches on Elysium: +10.6 points a trade against a coin flip (−27.8 to +54.0), 15 trades. A rule, never a coin; not financial advice.
- EXCITESDo not chaseOn launches on Elysium: +12.3 points a trade against a coin flip (+1.3 to +24.9), 84 trades. A rule, never a coin; not financial advice.
- UNRESOLVEDFollow the volumeOn launches on Elysium: +8.8 points a trade against a coin flip (−4.3 to +23.4), 74 trades. A rule, never a coin; not financial advice.
- EXCITESQuick flipOn launches on Elysium: +6.7 points a trade against a coin flip (+1.1 to +12.9), 87 trades. A rule, never a coin; not financial advice.
- EXCITESBuy in the first minuteOn launches on Elysium: +3.4 points a trade against a coin flip (+0.9 to +6.6), 107 trades. A rule, never a coin; not financial advice.
Where OPSIN was wrong
- Cautious risk managerIt predicted INERT; the verdict was INHIBITS −15.4, on Agent safety, Haiku 5.5 with thinking off. Notebook #035
- Double-checkIt predicted UNRESOLVED; the verdict was EXCITES +13.8, on DeFi maths, Haiku 5.5 with thinking off. Notebook #044
- Calculator toolIt predicted INERT; the verdict was EXCITES +11.9, on Markets maths, Haiku 5.5 with thinking off. Notebook #039
What it foresaw
- Foresight1.47 bits of surprise per verdict on 81 parts, committed before each beacon; the forecast it writes scored 1.60 on the same parts. An even guess scores 2.00; lower is better.
- A penalty threatIt did not see this coming: INERT −0.3 on Markets maths, Haiku 5.5 with thinking off, which its foresight gave 2%. Notebook #067
- Units on every amountIt did not see this coming: INHIBITS −5.0 on Markets maths, Haiku 5.5 with thinking off, which its foresight gave 9%. Notebook #067
The lab’s own checks
- PlaceboA do-nothing line came out clear in 8 of 80 notebooks.
- ControlsThe unchanged agent differed from itself in 3 of 80 notebooks.
- PredictionsOPSIN called 145 of 346 verdicts before the beacon drew the tasks.
What it tests next
- Verify exact arithmeticDeFi maths. Verdict 505 stopped early at 96 pairs with EXCITES +15.1 in Notebook #082; measured again at full size, with no early stop
- High-precision exp and divisionDeFi maths. Verdict 506 stopped early at 192 pairs with EXCITES +10.4 in Notebook #082; measured again at full size, with no early stop
- Answer onlyDeFi maths. Held at −34.4 on Haiku 5.5 with thinking off; does it matter on Sonnet 5.5?
Signed by the lab’s key 0xFf8E1E8ebF21064Ef781B75CdC8DFbd969399EbD. Digest a84e3b80f610870b88a5ba22c90ff46d121ee3e81ffbb490fa1c3741855e0224. The signed JSON. Every issue.